Declining Threshold for Hypoxia in the Gulf of Mexico - Environmental

Monitoring population-level responses of marine mammals to human activities. Erica Fleishman , Daniel P. Costa , John Harwood , Scott Kraus , David Mo...
7 downloads 11 Views 503KB Size
Environ. Sci. Technol. 2005, 39, 716-723

Declining Threshold for Hypoxia in the Gulf of Mexico C R A I G A . S T O W , * ,† S O N G S . Q I A N , ‡ A N D J. KEVIN CRAIG‡ Department of Environmental Health Sciences, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina 29208, and Nicholas School of the Environment and Earth Sciences, Duke University Marine Laboratory, Beaufort, North Carolina 28516-9721

The northwestern Gulf of Mexico shelf has been nicknamed “The Dead Zone” due to annual summertime (MaySeptember) bottom-water hypoxia (dissolved oxygen e2 mg L-1) that can be extensive (>20 000 km2) and last for several months. Hypoxia has been attributed to eutrophication caused by increasing nitrogen loads, although directly linking hypoxia to nitrogen is difficult. While the areal extent of hypoxia has been shown to increase with Mississippi River flow, it is unclear whether this increase results from enhanced vertical water-column stratification or from eutrophication caused by river-borne nutrients. Disentangling the relative contributions of eutrophication versus stratification has important management consequences. Our analysis indicates that the top:bottom salinity difference is an important predictor of hypoxia, exhibiting a threshold, where the probability of hypoxia increases rapidly, at approximately 4.1 ppt. Using a Bayesian change-point model, we show that this stratification threshold decreased from 1982 to 2002, indicating the degree of stratification needed to induce hypoxia has gone down. Although this declining threshold does not link hypoxia and nitrogen, it does implicate a long-term factor transcending yearly flowinduced stratification differences. Concurrently, we show that surface temperature increased, while surface dissolved oxygen decreased, suggesting that factors in addition to nitrogen may be influencing the incidence of hypoxia in the bottom water.

Introduction Coastal hypoxia is an increasingly well-documented global phenomenon (1-3). Hypoxia (dissolved oxygen e2 mg L-1) occurs when the oxygen depletion rate in the bottom of the water column exceeds the supply rate for an extended time. In coastal areas, this imbalance typically arises from a densityinduced stratification, when buoyant freshwater and dense ocean water form nearly distinct vertical layers that inhibit surface to bottom turbulent mixing. Although hypoxia can occur naturally, the incidence of hypoxia is believed to be increasing due to increasing cultural eutrophication (4-6). Coastal eutrophication is generally attributed to enhanced nitrogen input (7), although seasonal phosphorus limitation may occur in some systems (8-11). The spatial and temporal extent of hypoxia can differ substantially among coastal ecosystems, from patchy oc* Corresponding author phone: (803)777-6634; fax: (803)777-3391; e-mail: [email protected]. † University of South Carolina. ‡ Duke University. 716

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 39, NO. 3, 2005

currences lasting hours to days (12, 13), to events over much larger scales. The annual occurrence of summertime (MaySeptember) hypoxia on the northwestern Gulf of Mexico continental shelf has received considerable attention because it can extend over a large area (>20 000 km2) and last for several months (14). Nutrient loading from the MississippiAtchafalaya River system has historically been associated with high standing stocks of several ecologically and economically important fishes and other upper trophic levels, leading to the description of this region as the “Fertile Crescent” (15, 16). However, more recently, nutrient overenrichment and associated hypoxia have led to considerable habitat loss for these species with potential consequences for production (17-20). Effective nutrient management in this system is a particular challenge because the watershed includes ∼41% of the continental United States and encompasses many human activities (21, 22). To assess management strategies for reducing hypoxia in the Gulf of Mexico, it is important to distinguish the anthropogenic and “natural” factors that cause hypoxia, and the temporal scales on which they operate. However, sorting out the relative importance of the physical, chemical, and biological precursors of hypoxia is difficult because these factors are often interactive and highly confounded. River flow (volume/time), which is directly related to nutrient loading (mass/time), also affects salinity, temperature, stratification, turbidity, and residence time, factors that all influence primary productivity. Additionally, the relationship between river flow and riverine nutrient concentration can be either positive or negative, or nonmonotonic (23). The net result can be that nutrient concentration and primary productivity are nonmonotonic with respect to river flow, and the response function of these characteristics may differ systematically along spatial gradients (24). In the Gulf of Mexico, it is of particular interest to differentiate the physical (e.g., stratification) and biological (e.g., nutrients and associated primary production) effects associated with high river flow. While the areal extent of hypoxia has been shown to increase during high Mississippi River flow (25), it is unclear whether this increase results principally from increased stratification or enhanced primary productivity induced concurrently by high nutrient loading. Distinguishing these effects is important as they have different implications for managing the hypoxia problem. Long-term monitoring data are a valuable resource to help disentangle some of these processes because they provide a set of natural experiments at the ecosystem scale (26). Additionally, longterm data offer a useful context for evaluating the utility and possible success of alternative management actions (27).

Materials and Methods To examine processes related to hypoxia in the Gulf of Mexico, we analyzed a long-term data set (1982-2002) consisting of physical and chemical measurements, collected and provided by the NOAA National Marine Fisheries Service Pascagoula laboratory from their annual Southeast Area Monitoring and Assessment Program (SEAMAP) summer survey. This survey consists of a combined hydrographic and bottom trawl survey conducted during June-July in the northwestern Gulf of Mexico (Mississippi-Texas/Mexico border) since 1982 (28). Surveys used a stratified random design with strata based on depth and geographic location. Each survey was conducted by the R/V Oregon II with supplementary sampling by two additional vessels, the R/V Pelican and the R/V Tommy Munro, using similar sampling protocols. No sampling was conducted inshore of 3.7 m due to depth limitations of the 10.1021/es049412o CCC: $30.25

 2005 American Chemical Society Published on Web 12/22/2004

logit (p[ij]) ) R[i] + β1[i](salinity difference[ij] - change point[i]) if salinity difference[ij] e change point[i] or

logit (p[ij]) ) R[i] + β2[i](salinity difference[ij] - change point[i]) if salinity difference[ij] > change point[i] and FIGURE 1. Map depicting sample sites over entire period. survey vessels. In most cases, surface and bottom measurements of temperature, salinity, and dissolved oxygen were taken from CTD (conductivity, temperature, depth) profiles at each station, although in other cases environmental measurements were taken with other gear or determined from laboratory analysis of water samples. From 1982 to 1986, a slightly different survey design was used with all trawling conducted at night (as opposed to approximately equal day and night sampling) at depths from 9.1 to 91.4 m (as opposed to 3.7-109.7 m). Additional details regarding vessel characteristics, survey design, and sampling procedures can be found in Eldridge (28), Craig et al. (19, 20), Nichols et al. (29, 30), and the annual SEAMAP atlases (31). We confined our analysis to measurements taken within a rectangle defined by 28-30° N latitude and 89.5-94° W longitude (Figure 1), an area approximately encompassing the distribution of hypoxia since systematic mapping began in the early-mid 1980s (14, 19). We also excluded stations where the depth was >50 m, because exploratory analysis revealed hypoxia to be almost nonexistent at these deeper stations. The number of stations available for this analysis ranged from 42 to 142 (average ) 112) per survey. We used a combination of frequentist and Bayesian models to examine relationships among the data. For models based on assumptions of linearity and normality, frequentist approaches often provide an adequate basis for inference, whereas Bayesian methods are well-suited for nonlinear, nonnormal models, or situations where probabilistic inference is useful. To look at changes over time and relationships with river flow, we estimated simple linear regression models and plotted generalized additive models (GAMs) (32) to highlight deviations from linearity. For each year, we used measurements taken during June and July as realizations of the response variable in the linear regression and generalized additive models. Flow effects were examined by using the average daily Mississippi River flow at Tarbert Landing from April-July for each year as the predictor variable in the models (data available from the New Orleans District U.S. Army Corps of Engineers: http://www.mvn.usace.army.mil/eng/edhd/ watercon.htm). To evaluate the best predictor variable for bottom water oxygen concentration, we examined the data using classification and regression tree (CART) models (33, 34) and GAMs. Candidate predictor variables included surface and bottom salinity, salinity difference, surface and bottom temperature, temperature difference, depth, and surface water oxygen concentration. These exploratory analyses indicated that salinity difference (bottom salinity concentration minus surface salinity concentration) was the best overall predictor of bottom water oxygen concentration and suggested the presence of a salinity difference threshold, where the probability of hypoxia increased rapidly. To model this threshold, we used a Bayesian hierarchical change point model of the form:

Y[ij] ∼ bernoulli(p[ij]) where Y[ij] ) 1 if DO e 2 mg L-1 and 0 if DO > 2 mg L-1, p[ij] is the probability of hypoxia, logit p[ij] ) log(p[ij]/(1 - p[ij])), and [ij] represents the jth observation in year i. The observed Y[ij] is modeled as a realization from a random Bernoulli process where p[ij], the probability of hypoxia, is a function of the salinity difference. This model is an extension to the simple change point regression model in Stephens (35) and Qian and Richardson (36). The model represents two lines that meet at the change point, where logit (p[ij]) ) R[i]. The line to the left has a slope of β1[i] and an intercept of R[i] - β1[i] (changepoint[i]), and the line to the right has a slope of β2[i] and an intercept of R[i] - β2[i] (changepoint[i]). β1[i] represents the increase in the logit of the probability of hypoxia below the change point, while β2[i] is the increase in the logit of the probability of hypoxia above the change point. Exploratory estimates of this model indicated that β1[i] was essentially unchanging with time and that β2[i] had a high posterior probability of equaling zero each year, so for the final model we held β1[i] the same across years (or simply β1) and β2[i] fixed at zero for all years. With β2[i] fixed at zero, the quantity {R[i] - β2[i] (changepoint[i])} is equal to R[i] and provides an estimate of the probability of hypoxia after crossing the salinity difference threshold.

Results The linear regression model results illustrate an important limitation of classical statistical inference known as Lindley’s paradox (37). With almost 2400 observations collected over 21 years, few estimated models were not “statistically significant” (Tables 1 and 2); thus, “significance” is not a helpful criterion for deciding the importance of the observed relationships in this analysis. This limitation of classical significance testing as a basis for inference will become increasingly apparent as current long-term and proposed large-scale environmental data collection efforts produce enormous amounts of data (38-41). Therefore, our interpretation of these results focuses on the effect size and direction (the slope), the relative variance resolved by the model (R2), and whether the GAMs confirm that the underlying linearity assumption is met. Flow Relationships. Mississippi River flow from 1982 to 2002 exhibits some periodicity with the 1980s being generally dryer, the 1990s generally wetter, and a return to low flows by the late 1990s. This general pattern is overlain by considerable year-to-year variability, and there is no pronounced net change over this time period (Figure 2). Flow effects on temperature are minimal (Figure 3). Regression lines are nearly coincident with the overall average for surface temperature, and the surface:bottom temperature difference. This lack of relationship is reflected in R2 values, indicating that less than 1% of the overall variance is resolved by the linear flow model (Table 1). VOL. 39, NO. 3, 2005 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

717

TABLE 1. Summary Statistics for the Linear Regression Models Examining River Flow Relationships flow models

n

β0 (SE)

β1 (SE)

MSE

model F

P-value

R2

surface temperature bottom temperature temperature difference surface salinity bottom salinity salinity difference surface DO bottom DO surface % DO saturation bottom % DO saturation surface DO saturation concentration bottom DO saturation concentration

2404 2396 2396 2388 2385 2370 2347 2406 2332 2382 2388 2382

29.2 (0.12) 24.7 (0.23) -4.5 (0.24) 34.8 (0.53) 36.4 (0.49) 1.84 (0.45) 6.48 (0.13) 4.4 (0.19) 102.2 (1.81) 65.4 (2.83) 6.32 (0.02) 6.78 (0.002)

0.01 (0.006) 0.05 (0.01) 0.04 (0.01) -0.50 (0.03) -0.22 (0.02) 0.27 (0.02) 0.02 (0.006) -0.02 (0.01) 0.09 (0.09) -0.29 (0.14) 0.02 (0.002) 0.002 (0.001)

2.23 8.0 8.23 41.0 34.5 28.9 2.41 5.5 475.8 1167.7 0.076 0.075

4.76 18.5 8.86 352.7 79.7 147 13.4 2.8 1.0 4.42 216.9 3.36

0.03